Personalized rehabilitation scheme recommendation system for movement cognitive impairment of old people

Abstract

The invention discloses a personalized rehabilitation scheme recommendation system for movement cognitive impairment of old people. The system comprises a data acquisition module which is used for collecting case data about a patient; the blood test data processing module is used for performing embedded representation on the blood test data of the patient; the chief complaint text processing module is used for performing embedded representation on the patient chief complaint text; the rehabilitation scheme acquisition module is used for characterizing a single rehabilitation measure, completing characterization of a composite rehabilitation scheme, predicting the rehabilitation measure to perform event classification based on patient personalization, simulating interaction of multiple rehabilitation measures and obtaining characterization of the rehabilitation scheme; and the rehabilitation scheme recommendation module is used for recommending a personalized rehabilitation scheme for the motor cognitive impairment of the old people by utilizing the trained rehabilitation measure scoring model. According to the method, the problems of slow convergence process and even over-fitting caused by lack of an embedding process for users and articles in a current recommendation algorithm are solved; and patient personalization is not considered in single rehabilitation measure recommendation.

Description

Technical Field

[0001] The present invention relates to the field of rehabilitation medicine, and in particular to a personalized rehabilitation program recommendation system for elderly people with motor cognitive disorders. Background Art

[0002] With the development of medical information technology and big data technology, the integrated platform for rehabilitation medical evaluation and intervention has gradually become a research hotspot in the medical field. Traditional rehabilitation methods mainly rely on the experience and expertise of doctors. Although this method is effective, it fails to fully consider the patient's personalized physical condition and only provides general cognitive impairment treatment. The rehabilitation program is not targeted enough, resulting in poor rehabilitation efficiency and effect (AI-based ability assessment and rehabilitation treatment recommendation system (CN202311358422.1) Method and system for describing and recommending the best rehabilitation plan in adaptive telemedicine (CN202110443973.2)). Summary of the invention

[0003] The main research object of this invention is the elderly with cognitive impairment. The core of developing an integrated platform is how to use the patient's medical records and related information to provide patients with personalized rehabilitation plans to improve the accuracy and efficiency of rehabilitation. Existing studies have shown that collaborative filtering (CF) and graph convolutional networks (GCN) play an important role in multi-objective combination recommendation.

[0004] In order to solve the defects in the prior art, the present invention combines the classic recommendation algorithm with the deep learning network, and provides a personalized recommendation system for rehabilitation programs for motor cognitive disorders in the elderly based on patient blood test data and complaints using collaborative filtering and graph convolutional networks. The system has the following advantages for recommending rehabilitation programs based on patient blood test data and complaints: First, the blood test data can be obtained through a single blood test. This inspection method is both fast and accurate, and can fully reflect the patient's health status and multiple physiological indicators in the body. The blood test data include but are not limited to key indicators such as glycosylated hemoglobin, low-density lipoprotein, and triglycerides, which provide reliable basic information for recommending rehabilitation programs. Secondly, the blood test data of patients with the same disease often reflect different personalized signs. These personalized signs not only reflect the commonality of the disease, but also reveal the differences between individual patients. Based on these personalized signs, a more accurate personalized rehabilitation recommendation program can be achieved, thereby improving the rehabilitation effect. Thirdly, while considering the blood test data, the rehabilitation program recommendation also combines multimodal information, which includes multiple data sources such as the patient's complaint text and medical records. By integrating multimodal information, the system can have a more comprehensive understanding of the patient's health status, thereby improving the accuracy and reliability of the recommendation. Finally, the rehabilitation recommendation system can comprehensively analyze the self-consistent relationship between compound rehabilitation exercises and the patient's physical condition, and predict the rehabilitation effects of different rehabilitation programs on individual patients based on this, avoiding potential adverse reactions. This rehabilitation treatment program that integrates multiple single rehabilitation exercises makes the rehabilitation program recommendation more scientific and safe, and shortens the patient's rehabilitation cycle.

[0005] The present invention combines the advantages of collaborative filtering and graph convolutional networks, and proposes a personalized recommendation system for rehabilitation programs for motor cognitive disorders in the elderly. The system first embeds the patient's blood test data through an autoencoder, and uses a long short-term memory (LSTM) network with an attention mechanism to embed the patient's chief complaint text. Subsequently, a convolutional neural network is used to characterize the rehabilitation exercise, and the interaction between the rehabilitation exercise and the patient's personalization is simulated by initializing the graph convolutional network, and finally the characterization of the rehabilitation program is obtained through information aggregation and transmission. Finally, a recommendation model with collaborative filtering and an MLP interaction layer is combined to generate a rehabilitation treatment recommendation score.

[0006] This system can not only improve the accuracy and rationality of rehabilitation plan recommendations, but also help doctors make more scientific decisions in certain complex situations, improve patients' rehabilitation effects and efficiency, and reduce medical risks.

[0007] The present invention focuses on the recommendation of rehabilitation programs for motor cognitive disorders based on the patient's blood test data and chief complaint. The present invention collects case data about the patient, including the patient's blood test data, the chief complaint text and the historical rehabilitation program as the training truth value; the patient's blood test data is embedded and represented by an autoencoder and an MSE loss function; the patient's chief complaint text is embedded and represented by an LSTM network with an attention mechanism; then a single rehabilitation program (a single rehabilitation exercise) is further characterized by one-hot encoding, and then a rehabilitation program (a combination of multiple single rehabilitation exercises) is characterized by CNN-RRI-GCN (Convolutional Neural Networks-Rehabilitation and Rehabilitation Interaction-Graph Convolutional Networks); collaborative filtering with an MLP interaction layer and BPRLoss loss function are used to train to obtain a recommended score for the rehabilitation treatment program. The present invention solves the problem that the current recommendation algorithm generally lacks an embedding process for users and items, resulting in a slow convergence process or even overfitting; and the problem that the recommendation of a single rehabilitation measure lacks consideration of patient personalization.

[0008] The purpose of the present invention is achieved by at least one of the following technical solutions.

[0009] A personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment, specifically including:

[0010] Data acquisition module: collects case data about patients, including blood test data, chief complaint text and historical rehabilitation plan;

[0011] Blood test data processing module: uses autoencoder and MSE loss function to embed patient blood test data;

[0012] Chief complaint text processing module: an LSTM network with an attention mechanism is used to embed the patient's chief complaint text;

[0013] Rehabilitation plan acquisition module: One-hot encoding is used to characterize a single rehabilitation measure, and then the convolutional neural network rehabilitation measure interaction graph convolutional neural network CNN-RRI-GCN (Convolutional Neural Network-Rehabilitation Interaction-Graph Convolutional Network) is used to complete the characterization of the composite rehabilitation plan. The convolutional neural network is used to predict the rehabilitation measures for patient-based personalized event classification, and the graph convolutional network is initialized based on this to simulate the interaction of multiple rehabilitation measures. The characterization of the rehabilitation plan is obtained through information aggregation, transmission and global node pooling;

[0014] Rehabilitation plan recommendation module: Collaborative filtering with an MLP interaction layer is used to construct a rehabilitation measure scoring model, and then the BPRLoss loss function is used to train the trained rehabilitation measure scoring model. Finally, the trained rehabilitation measure scoring model is used to recommend personalized rehabilitation plans for motor cognitive disorders in the elderly.

[0015] Furthermore, in the data processing module, in order to apply the blood test data research, the patients are patients with heart disease, diabetes, hepatitis B disease and the above combined diseases;

[0016] The blood test data items include: glycosylated hemoglobin (Alc), low-density lipoprotein (Ldl), hepatitis B virus core antibody (Hbca), hepatitis B virus surface antibody (Hbsa), triglycerides (Tg), platelets (Plt), red blood cells (Rbc), cholesterol (Lym), hepatitis B e antigen (Hbeag), hepatitis B surface antigen (Hbsag), hepatitis B e antibody (Hbvea) and hemoglobin (Hbg).

[0017] Furthermore, in the blood test data processing module, an autoencoder is used to compress the patient's blood test data. The autoencoder includes an encoder and a decoder. The encoder and the decoder are both a three-layer MLP. The blood test data is used as an input vector x and mapped to an embedding vector u1 through the encoder, as follows:

[0018] H e1 =g(W e1 ·x+b e1 ) (1)

[0019] H e2 =g(W e2 ·H e1 +b e2 ) (2)

[0020] u1=W ep ·H e2 +b ep (3)

[0021] in, Represents a space of n-dimensional column vectors, that is, the input vector x belongs to the n-dimensional real number space; represents a space of p-dimensional column vectors, that is, the embedding vector u1 belongs to the p-dimensional real space; n represents the dimension of the input vector x, which is equal to the number of features of the blood test data; p is the dimension of the embedding vector u1, which represents the dimension after data compression; g is the ReLu activation function; W e1 , W e2 and W ep are weight matrices, W e1is the first hidden layer H used to map the input vector x to the encoder e1 The weight matrix W e2 is used to transform the first hidden layer H e1 Mapped to the second hidden layer H in the encoder e2 The weight matrix W ep is used to convert the second hidden layer H in the encoder e2 The weight matrix mapped to the embedding vector u1; b e1 、b e2 and b ep Respectively represent the bias vectors of the first hidden layer, the second hidden layer, and the third hidden layer in the autoencoder;

[0022] Map the embedding vector u1 back to the reconstructed input through the decoder The details are as follows:

[0023] H d1 =f(W d1 u1+b d1 ) (4)

[0024] H d2 =f(W d2 ·H d1 +b d2 ) (5)

[0025]

[0026] Where f is the ReLu activation function; W d1 , W d2 and W dx are weight matrices, W d1 is the first decoding hidden layer H used to map the embedding vector u1 into the decoder d1 The weight matrix W d2 is used to decode the first hidden layer H d1 Mapped to the second decoding hidden layer H in the decoder d2 The weight matrix W dx is used to decode the second hidden layer H d2 Mapped to the reconstructed output vector The weight matrix of d1 、b d2 and b dx They respectively represent the bias vectors of the first decoding hidden layer, the second decoding hidden layer, and the third decoding hidden layer in the decoder.

[0027] Furthermore, in the blood test data processing module, the mean square error (MSE) is used as the loss function to calculate the output vector reconstructed by the autoencoder The difference between the input vector x and the loss function The expression is:

[0028]

[0029] Among them, x i represents the i-th component in the input vector x, representing the specific value of the i-th blood test feature in the blood test data; Represents the reconstructed output vector The i-th component in is the data decoded from the encoder, which is consistent with the i-th component x in the input vector x. i corresponding;

[0030] Set the convergence condition if If it is less than 0.001, stop the iteration; otherwise, perform the following steps to update the parameters:

[0031]

[0032] m t =β1m t-1 +(1-β1)g t (9)

[0033]

[0034] Among them, t represents the number of iterations, which is used to identify the current iteration step during the optimization process; g t represents the gradient at the tth iteration; m t and v t Respectively represent the estimated values ​​of the first-order moment and the second-order moment at the tth iteration; W t and b t Respectively represent the weight matrix and bias vector at the tth iteration; represents the bias-corrected second-order moment estimate;

[0035] After iterative calculation and convergence conditions are met, the output u1 in the autoencoder is the patient's blood test data represented by the embedding.

[0036] Furthermore, in the chief complaint text processing module, the chief complaint text of the patient's conversation with the doctor during the consultation is preprocessed and filled with empty characters to form a fixed-length text note; the chief complaint text note = (w1, w2, ..., w 200 ) for each word w p Mapped to a vector The vectors y1, y2, y3, ..., y corresponding to the 200 words 200 The text matrix encoding Y can be composed to complete the word embedding of the entire text, as follows:

[0037] Word2 Vec(note)→Y (14)

[0038] Input the matrix Y into the LSTM network, and the cell state update calculation process is as follows:

[0039] f t =σ(W f ·[h t-1 ,y t ]+b f ) (15)

[0040] i t =σ(W p ·[h t-1 ,y t ]+b i ) (16)

[0041]

[0042] o t =σ(W o ·[h t-1 ,y t ]+b o ) (19)

[0043] h t =o t *tanh(C t ) (20)

[0044] Among them, W f , W p , W C and W o are weight matrices, W f Represents the weight matrix of the forget gate, W p represents the weight matrix of the input gate, W C The weight matrix representing the candidate memory cell state, W o represents the weight matrix of the output gate; b f , b i , b C and b o are bias vectors, b f represents the bias vector of the forget gate, b i represents the bias vector of the input gate, b C The bias vector representing the candidate memory cell state, b o represents the bias vector of the output gate; h t is the hidden state of the LSTM network at time step t; C t is the unit state; σ is the sigmoid activation function; * represents the hardmand product;

[0045] The hidden state matrix of the above LSTM network is H = [h1, h2, ..., h m ,...h 200 ] as input, calculate the weight matrix A, the elements of the weight matrix A Represents the importance of the hidden state of the LSTM network at time step m to time step n; the weight matrix A is expressed in matrix form:

[0046] A=H·H T (twenty one)

[0047] The elements in the weight matrix A are row normalized by softmax:

[0048] A′=Softmax(A) (22)

[0049] Element-wise weighting; for each time step m, the hidden state h of the LSTM network at that time step m The corresponding row a in the normalized weight matrix A′ m Perform element weighting, each hidden state h m will be converted into a weighted hidden state:

[0050]

[0051] The matrix is ​​represented as V = A'H;

[0052] Finally, pooling is performed to obtain the embedding representation u2 of the patient's complaint text:

[0053]

[0054] Among them, V n Represents the nth row of the matrix V.

[0055] Furthermore, in the rehabilitation program acquisition module, the classification and prediction of RRI are studied based on the classification characteristics of rehabilitation measures; the classification of rehabilitation measures includes cognitive behavioral therapy, memory training, psychotherapy, and gait training;

[0056] However, considering that the above-mentioned rehabilitation exercise classification feature information needs to be represented by a binary feature vector with a large dimension, the number of categories far exceeds the number of historical rehabilitation measures used in the cases, so its sparsity and dimensionality characteristics are not suitable as the basis for the representation of rehabilitation measures;

[0057] In contrast, it is more reasonable to use one-hot encoding to represent rehabilitation measures, that is, to use a vector d k to characterize a single rehabilitation measure, where d k =(0, 0, ..., 1, ..., 0), only the kth element is 1, and the rest are 0; dk Represents the one-hot encoding of the kth rehabilitation measure, where the value of k ranges from 1 to N, and N is the number of different rehabilitation measures found from historical rehabilitation plans;

[0058] Predicting RRI using CNN: Combine the classification features of rehabilitation exercises, calculate the similarity matrix, and select the corresponding two row vectors to form a feature matrix of rehabilitation measure pairs as the input of the CNN model; then use the CNN model to predict the probability of occurrence of various RRI events. The specific calculation method is as follows:

[0059] For one of the measures, s a ,s b are the binary feature vectors of the a-th rehabilitation measure and the b-th rehabilitation measure regarding cognitive behavioral therapy, Represents 116 evaluation dimensions of cognitive behavior; calculates the Jaccard similarity coefficient of these two rehabilitation measures at the cognitive behavioral therapy level, expressed as follows:

[0060]

[0061] Among them, M 11 Represented in vector s a and b In the example, the corresponding element value is the number of subscripts of 1, M 01 Represented in vector s a and b In the example, the corresponding element values ​​are the number of subscripts of 0 and 1 respectively; M 10 Represented in vector s a and b The corresponding element values ​​are the number of subscripts of 1 and 0 respectively; the Jaccard similarity coefficient S a，b The value range of is [0, 1];

[0062] Similarly, the similarity coefficient of any two rehabilitation measures at the classification level is calculated. N represents the number of different rehabilitation measures found from the historical rehabilitation programs. Then, an N×N similarity evaluation matrix S is constructed. The element value of the ath row or bth column in the similarity evaluation matrix S is S a，b , represents the similarity coefficient between the a-th rehabilitation measure and the b-th rehabilitation measure at the cognitive behavioral therapy level, and S a，b =S b，a ; The element of the ath row or the ath column in the similarity evaluation matrix S represents the similarity feature vector of the ath rehabilitation measure and other rehabilitation measures at the classification level of the measure;

[0063] According to this similarity matrix construction method, the similarity matrix between rehabilitation measures at all measure classification levels (cognitive behavioral therapy, memory training, psychotherapy, gait training) can be constructed in the same way, denoted as cognitive behavioral therapy similarity matrix C, memory training similarity matrix M, psychotherapy similarity matrix P, gait training similarity matrix G;

[0064] For any one of the rehabilitation measures set, assuming it is the ath rehabilitation measure, there is a row vector in the four similarity matrices CM, P, and G, which is the similarity relationship feature vector between the ath rehabilitation measure and other rehabilitation measures at different measure classification levels; these four similar feature vectors are concatenated column by column to form a 4×N matrix I, where N is the number of different rehabilitation measures found from the historical rehabilitation plans, and I is the comprehensive similarity matrix of the ath rehabilitation measure and other rehabilitation measures at all levels;

[0065] A rehabilitation measure interaction relationship prediction network including multiple convolutional layers and fully connected layers is used to predict the interaction relationship between rehabilitation measures; the comprehensive similarity matrix I is input into the rehabilitation measure interaction relationship prediction network, and the features of the subspace are extracted from the input tensor; three fully connected layers are set in the rehabilitation measure interaction relationship prediction network, according to the 65 categories of RRI events divided in RRI-DMI (Rehabilitation and Rehabilitation Interaction-DynamicMeasure Index), and the interaction relationship between rehabilitation measures (A, B) can be divided into five types: A→B synergy, B→A synergy, no obvious interaction, A→B antagonism, B→A antagonism, where A and B correspond to the aforementioned i-th rehabilitation measure and j-th measure respectively; therefore, the number of neurons in the three fully connected layers is set to 256, 65 and 5 respectively; the first two fully connected layers use linear activation functions, and the last layer uses sigmoid activation function for normalized probability prediction. The five interaction relationships between rehabilitation measures can be predicted through the rehabilitation measure interaction relationship prediction network.

[0066] Furthermore, in the rehabilitation plan acquisition module, the training of the rehabilitation measure interaction relationship prediction network uses the Adam optimization algorithm, with an initial learning rate η = 0.001, and the loss function Loss is:

[0067]

[0068] Among them, y q The RRI truth value of the qth interaction relationship, 0 means there is no RRI, 1 means there is RRI; p qIt represents the predicted probability of the five interaction relationships, and the result with the largest probability is selected as the RRI prediction result;

[0069] The GCN edge weights for the interaction and characterization of multiple rehabilitation measures included in the rehabilitation program are initialized based on the RRI events predicted by CNN, as follows:

[0070] Get a case from the data set, record the rehabilitation plan used in the case as R, and represent the rehabilitation plan R with a weighted graph; for each rehabilitation measure in the rehabilitation plan R, initialize a node in the graph to represent it; for any two rehabilitation measures in the rehabilitation plan R, if the RRI of the two rehabilitation measures predicted by CNN exists, generate a directed edge between the corresponding two nodes in the graph structure with a weight attached:

[0071]

[0072] Among them, v and u represent two different rehabilitation measures in the rehabilitation program R, e vu represents the impact of rehabilitation measure v on u; considering the rehabilitation measure d u and d v The effects and extent of the interactions will also vary from patient to patient, so a mask model is used to analyze the personalized impact of rehabilitation measures on patients;

[0073] The mask model is an MLP, which receives the patient embedding vector u. The embedding vector is constructed by concatenating the two feature vectors obtained by the blood test data processing module and the chief complaint text processing module. The concatenated vector u is used to represent the patient's characteristics in a personalized way. After processing, a vector consistent with the representation dimension of a single rehabilitation measure is output, called a mask vector, and the mask vector is element-wise multiplied with the rehabilitation measure representation to obtain The details are as follows:

[0074]

[0075] Among them: the mask model MLP includes a hidden layer; the activation function uses the ReLu activation function, d u and d v They represent the initial representation vectors of rehabilitation measures u and v, respectively, obtained by one-hot encoding; and represents the representation vector of rehabilitation measures adjusted by the patient-individualized impact mask;

[0076] Calculated and After that, they are concatenated and input into a pre-trained MLP model. and The concatenated vectors are sequentially reduced in dimension, linearly mapped, and activated to calculate an influence factor c for updating the edge weights. vu , as follows:

[0077]

[0078] Among them, c vu represents the impact factor, reflecting the degree of influence of the interaction between rehabilitation measures v and u on a specific patient;

[0079] The impact factor c obtained through the above steps vu Includes information on the extent to which the interaction between the two single rehabilitation measures affects the specific individual patient, represents the updated edge weight, which includes the interaction strength between the rehabilitation measures after the patient's personalized factors; therefore, the edge weight can be updated accordingly:

[0080]

[0081] After completing the initialization of nodes and edges, use the GCN training method for calculation:

[0082] GCN summarizes and learns node representations by recursively aggregating and transforming the feature vectors of adjacent nodes; each layer update involves message passing, message aggregation, and node representation update, and its mathematical expression is:

[0083]

[0084] Among them, v and u represent two different rehabilitation measures in the rehabilitation program, representing two neighboring nodes in the graph, and there is a directed edge from u to v; represents the message vector transmitted from node u to node v; is the representation of node v in the lth layer of GCN; represents the weight of the directed edge from u to v; N(v) represents the neighborhood of node v, from which node v collects information to update its aggregate message Initialized to the corresponding rehabilitation measure representation d u , d v ;

[0085] The following calculation formula is used to obtain the embedded representation g of the rehabilitation program:

[0086]

[0087] Where: σ is the ReLu activation function.

[0088] Furthermore, in the rehabilitation program recommendation module, the embedding results u1 and u2 obtained according to formula (3) and formula (24) are concatenated to obtain the complete embedding u representing the lth patient. l , and characterization of rehabilitation measures k ,u l represents the personalized embedding vector of the th patient, which combines the embedding information of the patient’s blood test data and the chief complaint text. k Represents the one-hot encoding embedding of the k-th rehabilitation measure, which is used to represent a single rehabilitation measure; the concatenated vector [u l |d k ] is mapped to values ​​through multiple layers of fully connected layers Represents the personalized information u based on the patient l As a rehabilitation measure k Rating:

[0089] H e1 =g(W e1 ·[u l | d k ]+b e1 ) (35)

[0090] H e2 =g(W e2 ·H e1 +b e2 ) (36)

[0091] H e3 =g(W e3 ·H e2 +b e3 ) (37)

[0092]

[0093] Among them, g is the activation function; W e1 , W e2 , W e3 and W eo are all weight matrices; b e1 , b e2 , b e3 , b eo is the bias vector; H e1 , H e2 and H e3 are the outputs of the hidden layers respectively.

[0094] Furthermore, in the rehabilitation program recommendation module, an MLP model is used as a rehabilitation measure scoring model;

[0095] Using BprLoss as the loss function, the rehabilitation measure scoring model is trained. The mathematical expression of the loss function L1 is:

[0096]

[0097] Where: σ represents the sigmoid activation function; P n represents the historical rehabilitation plan used by the nth case in the data set; j and k represent a single rehabilitation measure in the rehabilitation measure set; is the L2 regularization used;

[0098] By minimizing the loss function, we can As large as possible, that is, for the patient in the nth record, the corresponding rehabilitation measure d j Compared with rehabilitation measures without a history of use k The higher the score, the more rehabilitation measures scoring model was obtained after the initial training, an MLP model that can perform personalized scoring of a single rehabilitation measure based on patient information;

[0099] The rehabilitation program calculated by formula (34) is embedded into g, replacing the single rehabilitation measure representation d in formula (35) and sent to the rehabilitation measure scoring model completed by the initial training. The loss function can be expressed as:

[0100]

[0101] Update iteration. If L2(W, b) is less than 0.01, stop iteration; otherwise, perform the following steps to update parameters:

[0102]

[0103] m t =β1m t-1 +(1-β1)g t (42)

[0104]

[0105] After iterative calculation and convergence conditions are met, the final rehabilitation measure scoring model is obtained, and the rehabilitation plan is scored based on the patient's personalization:

[0106] H e1 =g(W1·[u i |g j ]+b1) (47)

[0107] H e2 =g(W2·H e1 +b2) (48)

[0108] H e3 =g(W3·H e2 +b3) (49)

[0109]

[0110] Furthermore, in the rehabilitation program recommendation module, the recommendation process is as follows:

[0111] The final trained rehabilitation measure scoring model outputs a score for the rehabilitation plan based on the input patient embedding (including blood test data and chief complaint text) and historical rehabilitation plan embedding; this score reflects the patient's preference for different rehabilitation measures, thereby achieving personalized rehabilitation plan recommendations.

[0112] Rehabilitation program recommendation is of great significance to alleviating the shortage of medical resources and improving the efficiency and quality of rehabilitation. The present invention combines the advantages of collaborative filtering and graph convolutional networks. The blood test data of elderly patients with cognitive impairment are embedded through an autoencoder, and the LSTM network with an attention mechanism is used to embed the patient's chief complaint text. A convolutional neural network is used to characterize a single rehabilitation measure, and the interaction between rehabilitation measures is simulated by initializing the graph convolutional network. Finally, the representation of the rehabilitation program is obtained through information aggregation and transmission. Combined with collaborative filtering and a recommendation model with an MLP interaction layer, a personalized score for the rehabilitation program is generated. A personalized recommendation method for rehabilitation programs for elderly people with motor cognitive impairment based on patient blood test data and chief complaint information is proposed.

[0113] Compared with the existing technical solutions, the advantages of the technical solution of the present invention are as follows: the recommendation of a single rehabilitation measure combination based on the patient's blood test data and chief complaint has the following advantages: first, the blood test data can be obtained through a single blood test; second, the blood test data of patients with the same disease reflect personalized physical signs, thereby realizing personalized recommendations; third, considering multimodal information, the recommendation accuracy is high; fourth, the interaction between multiple rehabilitation measures can be analyzed.

[0114] In order to solve the above problems, the present invention mainly focuses on the software part in terms of construction: the MVC software structure is used to build the software framework, and Qt and Python are used to write the system's visual interface and the specific operating logic of each module to meet the system usage requirements.

[0115] (1) In order to embed patient information, a blood test data processing module and a chief complaint text processing module are proposed. The blood test data processing module learns the compressed representation of patient numerical data in an unsupervised learning manner by minimizing the reconstruction error, and uses an autoencoder and MSE loss function to embed the patient's blood test data; while the chief complaint text processing module uses an LSTM network with an attention mechanism to embed the patient's chief complaint text, that is, it controls the flow of information through a gating mechanism, captures the semantic relationship in the patient's chief complaint text, and focuses on the typical features of the patient's chief complaint text through an attention mechanism. It then processes the complex structure of the patient's multi-dimensional information and captures its features to obtain the patient's embedded representation.

[0116] (2) In order to characterize a single rehabilitation measure, a rehabilitation plan acquisition module based on a multi-layer convolutional network and a message aggregation mechanism is proposed. First, the single rehabilitation measure is binary encoded according to the binary descriptor, and the Jaccard similarity matrix is ​​calculated; then the feature vector is sent to the multi-layer convolutional network to capture the local features of the rehabilitation measure pair and predict the classification results of the interaction events of the rehabilitation measure pair; finally, the graph convolutional network is initialized based on the result, and the complex interaction of the rehabilitation plan is simulated through information transmission, node update and other processes. Finally, the node information is globally pooled to obtain the embedded representation of the rehabilitation plan.

[0117] (3) Finally, a rehabilitation plan recommendation module based on patient information is proposed. This model is based on collaborative filtering. The patient embedding and single rehabilitation measure representation are sent to the multi-layer perceptron to achieve high-order feature interaction and simulate the interaction effect between the patient and the single rehabilitation measure. Then, the single rehabilitation measure is divided into positive samples and negative samples according to whether there is an interaction history with the individual patient in the data set. The Bayesian personalized ranking loss function is calculated to make the personalized score of the positive sample single rehabilitation measure combination higher than that of the negative sample. The trained multi-layer perceptron can realize personalized recommendation of single rehabilitation measures. Finally, the patient embedding and rehabilitation plan embedding are sent to the model for retraining to obtain a personalized rehabilitation plan recommendation model based on patient information. BRIEF DESCRIPTION OF THE DRAWINGS

[0118] Figure 1 This is a workflow diagram of a personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment in an embodiment of the present invention;

[0119] Figure 2 This is a data processing structure diagram of the rehabilitation measure scoring model in an embodiment of the present invention;

[0120] Figure 3 This is a schematic diagram of the structure of a blood test data processing module in an embodiment of the present invention;

[0121] Figure 4 This is a structural diagram of a main complaint text processing module in an embodiment of the present invention;

[0122] Figure 5 4 is a flowchart of the rehabilitation program acquisition module in an embodiment of the present invention. DETAILED DESCRIPTION

[0123] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the specific implementation of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0124] Example:

[0125] A personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment, such as Figure 1 As shown, specifically including:

[0126] Data acquisition module: collects case data about patients, including blood test data, chief complaint text and historical rehabilitation plan;

[0127] Blood test data processing module: uses autoencoder and MSE loss function to embed patient blood test data;

[0128] Chief complaint text processing module: an LSTM network with an attention mechanism is used to embed the patient's chief complaint text;

[0129] Rehabilitation plan acquisition module: One-hot encoding is used to characterize a single rehabilitation measure, and then the convolutional neural network rehabilitation measure interaction graph convolutional neural network CNN-RRI-GCN (Convolutional Neural Network-Rehabilitation Interaction-Graph Convolutional Network) is used to complete the characterization of the composite rehabilitation plan. The convolutional neural network is used to predict the rehabilitation measures for patient-based personalized event classification, and the graph convolutional network is initialized based on this to simulate the interaction of multiple rehabilitation measures. The characterization of the rehabilitation plan is obtained through information aggregation, transmission and global node pooling;

[0130] Rehabilitation program recommended modules: Figure 2 As shown in the figure, collaborative filtering with an MLP interaction layer is used to construct a rehabilitation measure scoring model, and then the BPRLoss loss function is used to train the trained rehabilitation measure scoring model. Finally, the trained rehabilitation measure scoring model is used to recommend personalized rehabilitation plans for motor cognitive impairment in the elderly.

[0131] Furthermore, in the data processing module, in order to apply the blood test data research, the patients are patients with heart disease, diabetes, hepatitis B disease and the above combined diseases;

[0132] The blood test data items include: glycosylated hemoglobin (Alc), low-density lipoprotein (Ldl), hepatitis B virus core antibody (Hbca), hepatitis B virus surface antibody (Hbsa), triglycerides (Tg), platelets (Plt), red blood cells (Rbc), cholesterol (Lym), hepatitis B e antigen (Hbeag), hepatitis B surface antigen (Hbsag), hepatitis B e antibody (Hbvea) and hemoglobin (Hbg).

[0133] Furthermore, in the blood test data processing module, if Figure 3 As shown, an autoencoder is used to compress the patient's blood test data. The autoencoder includes an encoder and a decoder. The encoder and the decoder are both a three-layer MLP. The blood test data is used as an input vector x and mapped to an embedding vector u1 through the encoder, as follows:

[0134] H e1 =g(W e1 ·x+b e1 ) (1)

[0135] H e2 =g(W e2 ·H e1 +b e2 ) (2)

[0136] u1=W ep ·H e2 +b ep (3)

[0137] in, Represents a space of n-dimensional column vectors, that is, the input vector x belongs to the n-dimensional real number space; represents a space of p-dimensional column vectors, that is, the embedding vector u1 belongs to the p-dimensional real space; n represents the dimension of the input vector x, which is equal to the number of features of the blood test data; p is the dimension of the embedding vector u1, which represents the dimension after data compression; g is the ReLu activation function; W e1 , W e2 and W ep are weight matrices, W e1 is the first hidden layer H used to map the input vector x to the encoder e1 The weight matrix W e2 is used to transform the first hidden layer H e1 Mapped to the second hidden layer H in the encoder e2 The weight matrix W ep is used to convert the second hidden layer H in the encoder e2 The weight matrix mapped to the embedding vector u1; b e1 、b e2 and b epRespectively represent the bias vectors of the first hidden layer, the second hidden layer, and the third hidden layer in the autoencoder;

[0138] Map the embedding vector u1 back to the reconstructed input through the decoder The details are as follows:

[0139] H d1 =f(W d1 u1+b d1 ) (4)

[0140] H d2 =f(W d2 ·H d1 +b d2 ) (5)

[0141]

[0142] Where f is the ReLu activation function; W d1 , W d2 and W dx are weight matrices, W d1 is the first decoding hidden layer H used to map the embedding vector u1 into the decoder d1 The weight matrix W d2 is used to decode the first hidden layer H d1 Mapped to the second decoding hidden layer H in the decoder d2 The weight matrix W dx is used to decode the second hidden layer H d2 Mapped to the reconstructed output vector The weight matrix of d1 , b d2 and b dx They respectively represent the bias vectors of the first decoding hidden layer, the second decoding hidden layer, and the third decoding hidden layer in the decoder.

[0143] In one embodiment, in the blood test data processing module, the mean square error (MSE) is used as the loss function to calculate the output vector reconstructed by the autoencoder. The difference between the input vector x and the loss function The expression is:

[0144]

[0145] Among them, x i represents the i-th component in the input vector x, representing the specific value of the i-th blood test feature in the blood test data; Represents the reconstructed output vector The i-th component in is the data decoded from the encoder, which is consistent with the i-th component x in the input vector x. icorresponding;

[0146] Set the convergence condition if If it is less than 0.001, stop the iteration; otherwise, perform the following steps to update the parameters:

[0147]

[0148] m t =β1m t-1 +(1-β1)g t (9)

[0149]

[0150] Among them, t represents the number of iterations, which is used to identify the current iteration step during the optimization process; g t represents the gradient at the tth iteration; m t and v t Respectively represent the estimated values ​​of the first-order moment and the second-order moment at the tth iteration; W t and b t Respectively represent the weight matrix and bias vector at the tth iteration; represents the bias-corrected second-order moment estimate;

[0151] After iterative calculation and convergence conditions are met, the output u1 in the autoencoder is the patient's blood test data represented by the embedding.

[0152] In one embodiment, Figure 4 As shown in the figure, in the chief complaint text processing module, the chief complaint text of the patient's conversation with the doctor during the consultation is preprocessed and filled with empty characters to form a fixed-length text note with a length of 200; Word2Vec is used to convert the chief complaint text note = (w1, w2, ..., w 200 ) for each word w p Mapped to a vector The vectors y1, y2, y3, ..., y corresponding to the 200 words 200 The text matrix encoding Y can be composed to complete the word embedding of the entire text, as follows:

[0153] Word2Vec(note)→Y (14)

[0154] Input the matrix Y into the LSTM network, and the cell state update calculation process is as follows:

[0155] f t =σ(W f ·[h t-1 ,y t ]+b f ) (15)

[0156] i t =σ(W p ·[h t-1 ,y t ]+b i ) (16)

[0157]

[0158] o t =σ(W o ·[h t-1 ,y t ]+b o ) (19)

[0159] h t =o t *tanh(C t ) (20)

[0160] Among them, W f , W p , W C and W o are weight matrices, W f Represents the weight matrix of the forget gate, W p represents the weight matrix of the input gate, W C The weight matrix representing the candidate memory cell state, W o represents the weight matrix of the output gate; b f , b i , b C and b o are bias vectors, b f represents the bias vector of the forget gate, b i represents the bias vector of the input gate, b C The bias vector representing the candidate memory cell state, b o represents the bias vector of the output gate; h t is the hidden state of the LSTM network at time step t; C t is the unit state; σ is the sigmoid activation function; * represents the hardmand product;

[0161] The hidden state matrix of the above LSTM network is H = [h1, h2, ..., h m ,...h 200 ] as input, calculate the weight matrix A, the elements of the weight matrix A Represents the importance of the hidden state of the LSTM network at time step m to time step n; the weight matrix A is expressed in matrix form:

[0162] A=H·H T(twenty one)

[0163] The elements in the weight matrix A are row normalized by softmax:

[0164] A′=Softmax(A) (22)

[0165] Element-wise weighting; for each time step m, the hidden state h of the LSTM network at that time step m The corresponding row a in the normalized weight matrix A′ m Perform element weighting, each hidden state h m will be converted into a weighted hidden state:

[0166]

[0167] The matrix is ​​represented as V = A'H;

[0168] Finally, pooling is performed to obtain the embedding representation u2 of the patient's complaint text:

[0169]

[0170] Among them, V n Represents the nth row of the matrix V.

[0171] In one embodiment, Figure 5 As shown, in the rehabilitation program acquisition module, RRI classification and prediction are studied according to the classification characteristics of rehabilitation measures; the classification of rehabilitation measures includes cognitive behavioral therapy, memory training, psychotherapy and gait training;

[0172] However, considering that the above-mentioned rehabilitation exercise classification feature information needs to be represented by a binary feature vector with a large dimension, the number of categories far exceeds the number of historical rehabilitation measures used in the cases, so its sparsity and dimensionality characteristics are not suitable as the basis for the representation of rehabilitation measures;

[0173] In contrast, it is more reasonable to use one-hot encoding to represent rehabilitation measures, that is, to use a vector d k to characterize a single rehabilitation measure, where d k =(0, 0, ..., 1, ..., 0), only the kth element is 1, and the rest are 0; d k Represents the one-hot encoding of the kth rehabilitation measure, where the value of k ranges from 1 to N, and N is the number of different rehabilitation measures found from historical rehabilitation plans;

[0174] Predicting RRI using CNN: Combine the classification features of rehabilitation exercises, calculate the similarity matrix, and select the corresponding two row vectors to form a feature matrix of rehabilitation measure pairs as the input of the CNN model; then use the CNN model to predict the probability of occurrence of various RRI events. The specific calculation method is as follows:

[0175] For one of the measures, s a ,s b are the binary feature vectors of the a-th rehabilitation measure and the b-th rehabilitation measure regarding cognitive behavioral therapy, Represents 116 evaluation dimensions of cognitive behavior; calculates the Jaccard similarity coefficient of these two rehabilitation measures at the cognitive behavioral therapy level, expressed as follows:

[0176]

[0177] Among them, M 11 Represented in vector s a and b In the example, the corresponding element value is the number of subscripts of 1, M 01 Represented in vector s a and b In the example, the corresponding element values ​​are the number of subscripts of 0 and 1 respectively; M 10 Represented in vector s a and b The corresponding element values ​​are the number of subscripts of 1 and 0 respectively; the Jaccard similarity coefficient S a，b The value range of is [0, 1];

[0178] Similarly, the similarity coefficient of any two rehabilitation measures at the classification level is calculated. N represents the number of different rehabilitation measures found from the historical rehabilitation programs. Then, an N×N similarity evaluation matrix S is constructed. The element value of the ath row or bth column in the similarity evaluation matrix S is S a，b , represents the similarity coefficient between the a-th rehabilitation measure and the b-th rehabilitation measure at the cognitive behavioral therapy level, and S a，b =S b，a ; The element of the ath row or the ath column in the similarity evaluation matrix S represents the similarity feature vector of the ath rehabilitation measure and other rehabilitation measures at the classification level of the measure;

[0179] According to this similarity matrix construction method, the similarity matrix between rehabilitation measures at all measure classification levels (cognitive behavioral therapy, memory training, psychotherapy, gait training) can be constructed in the same way, denoted as cognitive behavioral therapy similarity matrix C, memory training similarity matrix M, psychotherapy similarity matrix P, gait training similarity matrix G;

[0180] For any one of the rehabilitation measures set, assuming it is the ath rehabilitation measure, there is a row vector in the four similarity matrices C, M, P, and G, which is the similarity relationship feature vector between the ath rehabilitation measure and other rehabilitation measures at different measure classification levels; these four similar feature vectors are concatenated column by column to form a 4×N matrix I, where N is the number of different rehabilitation measures found from the historical rehabilitation plans, and I is the comprehensive similarity matrix of the ath rehabilitation measure and other rehabilitation measures at all levels;

[0181] In one embodiment, a rehabilitation measure interaction relationship prediction network including multiple convolutional layers and fully connected layers is used to predict the interaction relationship between rehabilitation measures; the comprehensive similarity matrix I is input into the rehabilitation measure interaction relationship prediction network, and the features of the subspace are extracted from the input tensor; three fully connected layers are set in the rehabilitation measure interaction relationship prediction network, according to the 65 categories of RRI events divided in RRI-DMI (Rehabilitation and Rehabilitation Interaction-Dynamic Measure Index), and the interaction relationship between the rehabilitation measure pair (A, B) can be divided into 5 types: A→B synergy, B→A synergy, no obvious interaction, A→B antagonism, B→A antagonism, where A and B correspond to the aforementioned i-th rehabilitation measure and j-th measure respectively; therefore, the number of neurons in the three fully connected layers is set to 256, 65 and 5 respectively; the first two fully connected layers use linear activation functions, and the last layer uses sigmoid activation function to perform normalized probability prediction, and the five interaction relationships between rehabilitation measures can be predicted through the rehabilitation measure interaction relationship prediction network.

[0182] Furthermore, in the rehabilitation plan acquisition module, the training of the rehabilitation measure interaction relationship prediction network uses the Adam optimization algorithm, with an initial learning rate η = 0.001, and the loss function Loss is:

[0183]

[0184] Among them, y q The RRI truth value of the qth interaction relationship, 0 means there is no RRI, 1 means there is RRI; p q It represents the predicted probability of the five interaction relationships, and the result with the largest probability is selected as the RRI prediction result;

[0185] The GCN edge weights for the interaction and characterization of multiple rehabilitation measures included in the rehabilitation program are initialized based on the RRI events predicted by CNN, as follows:

[0186] Get a case from the data set, record the rehabilitation plan used in the case as R, and represent the rehabilitation plan R with a weighted graph; for each rehabilitation measure in the rehabilitation plan R, initialize a node in the graph to represent it; for any two rehabilitation measures in the rehabilitation plan R, if the RRI of the two rehabilitation measures predicted by CNN exists, generate a directed edge between the corresponding two nodes in the graph structure with a weight attached:

[0187]

[0188] Among them, v and u represent two different rehabilitation measures in the rehabilitation program R, e vu represents the impact of rehabilitation measure v on u; considering the rehabilitation measure d u and d v The effects and extent of the interactions will also vary from patient to patient, so a mask model is used to analyze the personalized impact of rehabilitation measures on patients;

[0189] The mask model is an MLP, which receives the patient embedding vector u. The embedding vector is constructed by concatenating the two feature vectors obtained by the blood test data processing module and the chief complaint text processing module. The concatenated vector u is used to represent the patient's characteristics in a personalized way. After processing, a vector consistent with the representation dimension of a single rehabilitation measure is output, called a mask vector, and the mask vector is element-wise multiplied with the rehabilitation measure representation to obtain The details are as follows:

[0190]

[0191] Among them: the mask model MLP includes a hidden layer; the activation function uses the ReLu activation function, d u and d v They represent the initial representation vectors of rehabilitation measures u and v, respectively, obtained by one-hot encoding; and represents the representation vector of rehabilitation measures adjusted by the patient-individualized impact mask;

[0192] Calculated and After that, they are concatenated and input into a pre-trained MLP model. and The concatenated vectors are sequentially reduced in dimension, linearly mapped, and activated to calculate an influence factor c for updating the edge weights. vu , as follows:

[0193]

[0194] Among them, cvu represents the impact factor, reflecting the degree of influence of the interaction between rehabilitation measures v and u on a specific patient;

[0195] The impact factor c obtained through the above steps vu Includes information on the extent to which the interaction between the two single rehabilitation measures affects the specific individual patient, represents the updated edge weight, which includes the interaction strength between the rehabilitation measures after the patient's personalized factors; therefore, the edge weight can be updated accordingly:

[0196]

[0197] After completing the initialization of nodes and edges, use the GCN training method for calculation:

[0198] GCN summarizes and learns node representations by recursively aggregating and transforming the feature vectors of adjacent nodes; each layer update involves message passing, message aggregation, and node representation update, and its mathematical expression is:

[0199]

[0200] Among them, v and u represent two different rehabilitation measures in the rehabilitation program, representing two neighboring nodes in the graph, and there is a directed edge from u to v; represents the message vector transmitted from node u to node v; is the representation of node v in the lth layer of GCN; represents the weight of the directed edge from u to v; Ν(v) represents the neighborhood of node v, from which node v collects information to update its aggregate message Initialized to the corresponding rehabilitation measure representation d u , d v ;

[0201] The following calculation formula is used to obtain the embedded representation g of the rehabilitation program:

[0202]

[0203] Where: σ is the ReLu activation function.

[0204] In one embodiment, in the rehabilitation program recommendation module, the embedding results u1 and u2 obtained according to formula (3) and formula (24) are concatenated to obtain the complete embedding u representing the lth patient. l , and characterization of rehabilitation measures k ,u l represents the personalized embedding vector of the lth patient, which combines the embedding information of the patient’s blood test data and the chief complaint text. kRepresents the one-hot encoding embedding of the k-th rehabilitation measure, which is used to represent a single rehabilitation measure; the concatenated vector [u l |d k ] is mapped to values ​​through multiple layers of fully connected layers Represents the personalized information u based on the patient l As a rehabilitation measure k Rating:

[0205] H e1 =g(W e1 ·[u l |d k ]+b e1 ) (35)

[0206] H e2 =g(W e2 ·H e1 +b e2 ) (36)

[0207] H e3 =g(W e3 ·H e2 +b e3 ) (37)

[0208]

[0209] Among them, g is the activation function; W e1 , W e2 , W e3 and W eo are all weight matrices; b e1 , b e2 , b e3 , b eo is the bias vector; H e1 , H e2 and H e3 are the outputs of the hidden layers respectively.

[0210] Furthermore, in the rehabilitation program recommendation module, Figure 2 As shown, an MLP model was used as the rehabilitation measure scoring model;

[0211] Using BprLoss as the loss function, the rehabilitation measure scoring model is trained. The mathematical expression of the loss function L1 is:

[0212]

[0213] Where: σ represents the sigmoid activation function; P n represents the historical rehabilitation plan used by the nth case in the data set; j and k represent a single rehabilitation measure in the rehabilitation measure set; is the L2 regularization used;

[0214] By minimizing the loss function, we can As large as possible, that is, for the patient in the nth record, the corresponding rehabilitation measure d j Compared with rehabilitation measures without a history of use k The higher the score, the more rehabilitation measures scoring model was obtained after the initial training, an MLP model that can perform personalized scoring of a single rehabilitation measure based on patient information;

[0215] The rehabilitation program calculated by formula (34) is embedded into g, replacing the single rehabilitation measure representation d in formula (35) and sent to the rehabilitation measure scoring model completed by the initial training. The loss function can be expressed as:

[0216]

[0217] Update iteration. If L2(W, b) is less than 0.01, stop iteration; otherwise, perform the following steps to update parameters:

[0218]

[0219] m t =β1m t-1 +(1-β1)g t (42)

[0220]

[0221]

[0222] After iterative calculation and convergence conditions are met, the final rehabilitation measure scoring model is obtained, and the rehabilitation plan is scored based on the patient's personalization:

[0223] H e1 =g(W1·[u i |g j ]+b1) (47)

[0224] H e2 =g(W2·H e1 +b2) (48)

[0225] H e3 =g(W3·H e2 +b3) (49)

[0226]

[0227] In one embodiment, in the rehabilitation program recommendation module, the recommendation process is as follows:

[0228] The final trained rehabilitation measure scoring model outputs a score for the rehabilitation plan based on the input patient embedding (including blood test data and chief complaint text) and historical rehabilitation plan embedding; this score reflects the patient's preference for different rehabilitation measures, thereby achieving personalized rehabilitation plan recommendations.

Claims

1. A personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment, characterized in that: Specifically include: Data acquisition module: collects case data about patients, including blood test data, chief complaint text and historical rehabilitation plan; Blood test data processing module: uses autoencoder and MSE loss function to embed patient blood test data; Chief complaint text processing module: an LSTM network with an attention mechanism is used to embed the patient's chief complaint text; Rehabilitation program acquisition module: using on e -hot encoding is used to characterize a single rehabilitation measure, and then a convolutional neural network rehabilitation measure interaction graph convolutional neural network CNN-RRI-GCN is used to complete the characterization of the composite rehabilitation plan. The convolutional neural network is used to predict rehabilitation measures for event classification based on patient personalization, and the graph convolutional network is initialized accordingly to simulate the interaction of multiple rehabilitation measures. The characterization of the rehabilitation plan is obtained through information aggregation, transmission and global node pooling; Rehabilitation plan recommendation module: Collaborative filtering with an MLP interaction layer is used to construct a rehabilitation measure scoring model, and then the BPRLoss loss function is used to train the trained rehabilitation measure scoring model. Finally, the trained rehabilitation measure scoring model is used to recommend personalized rehabilitation plans for motor cognitive disorders in the elderly.

2. According to claim 1, a personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment is characterized in that: In the data processing module, the patient is a patient with heart disease, diabetes, hepatitis B disease, or the above-mentioned combined diseases; Blood test data items include: glycated hemoglobin, low-density lipoprotein, hepatitis B virus core antibody, hepatitis B virus surface antibody, triglycerides, platelets, red blood cells, cholesterol, hepatitis B e antigen, hepatitis B surface antigen, hepatitis B e antibody and hemoglobin.

3. According to claim 1, a personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment is characterized in that: In the blood test data processing module, an autoencoder is used to compress the patient's blood test data. The autoencoder includes an encoder and a decoder. The encoder and the decoder are both a three-layer MLP. The blood test data is used as the input vector x and mapped to the embedding vector u1 through the encoder, as follows: H e1 =g(W e1 ·x+b e1 ) (1) H e2 =g(W e2 ·H e1 +b e2 ) (2) u1=W ep ·H e2 +b ep (3) in, Represents a space of n-dimensional column vectors, that is, the input vector x belongs to the n-dimensional real number space; represents a space of p-dimensional column vectors, that is, the embedding vector u1 belongs to the p-dimensional real space; n represents the dimension of the input vector x, which is equal to the number of features of the blood test data; p is the dimension of the embedding vector u1, which represents the dimension after data compression; g is the ReLu activation function; W e1 , W e2 and W ep are weight matrices, W e1 is the first hidden layer H used to map the input vector x to the encoder e1 The weight matrix W e2 is used to transform the first hidden layer H e1 Mapped to the second hidden layer H in the encoder e2 The weight matrix W ep is used to convert the second hidden layer H in the encoder e2 The weight matrix mapped to the embedding vector u1; b e1 , b e2 and b ep Respectively represent the bias vectors of the first hidden layer, the second hidden layer, and the third hidden layer in the autoencoder; Map the embedding vector u1 back to the reconstructed input through the decoder The details are as follows: H d1 =f(W d1 ·u1+b d1 ) (4) H d2 =f(W d2 ·H d1 +b d2 ) (5) Where f is the ReLu activation function; W d1 , W d2 and W dx are weight matrices, W d1 is the first decoding hidden layer H used to map the embedding vector u1 into the decoder d1 The weight matrix W d2 is used to decode the first hidden layer H d1 Mapped to the second decoding hidden layer H in the decoder d2 The weight matrix W dx is used to decode the second hidden layer H d2 Mapped to the reconstructed output vector The weight matrix of d1 , b d2 and b dx They respectively represent the bias vectors of the first decoding hidden layer, the second decoding hidden layer, and the third decoding hidden layer in the decoder.

4. A personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment according to claim 3, characterized in that: In the blood test data processing module, the mean square error (MSE) is used as the loss function to calculate the output vector reconstructed by the autoencoder The difference between the input vector x and the loss function The expression is: Among them, x i represents the i-th component in the input vector x, representing the specific value of the i-th blood test feature in the blood test data; Represents the reconstructed output vector The i-th component in is the data decoded from the encoder, which is consistent with the i-th component x in the input vector x. i corresponding; Set the convergence condition if If it is less than 0.001, stop the iteration; otherwise, perform the following steps to update the parameters: m t =β1m t-1 +(1-β1)g t (9) Among them, t represents the number of iterations, which is used to identify the current iteration step during the optimization process; g t represents the gradient at the tth iteration; m t and v t Respectively represent the estimated values ​​of the first-order moment and the second-order moment at the tth iteration; W t and b t Respectively represent the weight matrix and bias vector at the tth iteration; represents the bias-corrected second-order moment estimate; After iterative calculation and convergence conditions are met, the output u1 in the autoencoder is the patient's blood test data represented by the embedding.

5. According to claim 1, a personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment is characterized in that: In the chief complaint text processing module, the chief complaint text of the patient's conversation with the doctor during the consultation is preprocessed and filled with empty characters to form a fixed-length text note; the chief complaint text note = (w1, w2, ..., w 200 ) p Mapping to a vector The vectors y1, y2, y3, ..., y corresponding to the 200 words 200 The text matrix encoding Y is composed to complete the word embedding of the entire text, as follows: Word2Vec(note)→Y (14) Input the matrix Y into the LSTM network, and the cell state update calculation process is as follows: f t =σ(W f ·[h t-1 ,y t ]+b f ) (15) i t =σ(W p ·[h t-1 ,y t ]+b i ) (16) the t =σ(W o ·[h t-1 ,y t ]+b o ) (19) h t =o t *tanh(C t ) (20) Among them, W f , W p , W C and W o are weight matrices, W f Represents the weight matrix of the forget gate, W p represents the weight matrix of the input gate, W C The weight matrix representing the candidate memory cell state, W o represents the weight matrix of the output gate; b f , b i , b C and b o are bias vectors, b f represents the bias vector of the forget gate, b i represents the bias vector of the input gate, b C The bias vector representing the candidate memory cell state, b o represents the bias vector of the output gate; h t is the hidden state of the LSTM network at time step t; C t is the unit state; σ is the sigmoid activation function; * represents the hardmand product; The hidden state matrix of the LSTM network is H = [h1, h2, ..., h m ,...h 200 ] as input, calculate the weight matrix A, the elements of the weight matrix A Represents the importance of the hidden state of the LSTM network at time step m to time step n; the weight matrix A is expressed in matrix form: A=H·H T (21) The elements in the weight matrix A are row normalized by softmax: A′=Softmax(A) (22) Element-wise weighting; for each time step m, the hidden state h of the LSTM network at that time step m The corresponding row a in the normalized weight matrix A′ m Perform element weighting, each hidden state h m will be converted into a weighted hidden state: The matrix is ​​represented as V = A'H; Finally, pooling is performed to obtain the embedding representation u2 of the patient's complaint text: Among them, V n Represents the nth row of the matrix V.

6. The personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment according to claim 1 is characterized in that: In the rehabilitation program acquisition module, the classification and prediction of RRI are studied based on the classification characteristics of rehabilitation measures; the classification of rehabilitation measures includes cognitive behavioral therapy, memory training, psychotherapy and gait training; With a vector d k to characterize a single rehabilitation measure, where d k =(0, 0, ..., 1, ..., 0), only the kth element is 1, and the rest are 0; d k Represents the one-hot encoding of the kth rehabilitation measure, where the value of k ranges from 1 to N, and N is the number of different rehabilitation measures found from historical rehabilitation plans; Predicting RRI using CNN: Combine the classification features of rehabilitation exercises, calculate the similarity matrix, and select the corresponding two row vectors to form a feature matrix of rehabilitation measure pairs as the input of the CNN model; then use the CNN model to predict the probability of occurrence of various RRI events. The specific calculation method is as follows: For one of the measures, s a ,s b are the binary feature vectors of the a-th rehabilitation measure and the b-th rehabilitation measure regarding cognitive behavioral therapy, Represents 116 evaluation dimensions of cognitive behavior; calculates the Jaccard similarity coefficient of these two rehabilitation measures at the cognitive behavioral therapy level, expressed as follows: Among them, M 11 Represented in vector s a and b In the example, the corresponding element value is the number of subscripts of 1, M 01 Represented in vector s a and b In the example, the corresponding element values ​​are the number of subscripts of 0 and 1 respectively; M 10 Represented in vector s a and b The corresponding element values ​​are the number of subscripts of 1 and 0 respectively; the Jaccard similarity coefficient S d，b The value range of is [0, 1]; Similarly, the similarity coefficient of any two rehabilitation measures at the classification level is calculated. N represents the number of different rehabilitation measures found from the historical rehabilitation programs. Then, an N×N similarity evaluation matrix S is constructed. The element value of the ath row or bth column in the similarity evaluation matrix S is S a，b , represents the similarity coefficient between the a-th rehabilitation measure and the b-th rehabilitation measure at the cognitive behavioral therapy level, and S a，b =S b,a ; The element of the ath row or the ath column in the similarity evaluation matrix S represents the similarity feature vector of the ath rehabilitation measure and other rehabilitation measures at the classification level of the measure; According to this similarity matrix construction method, similarity matrices between rehabilitation measures at all measure classification levels were constructed in the same way, denoted as cognitive behavioral therapy similarity matrix C, memory training similarity matrix M, psychotherapy similarity matrix P, and gait training similarity matrix G; For any one of the rehabilitation measures set, assuming it is the ath rehabilitation measure, there is a row vector in the four similarity matrices C, M, P, and G, which is the similarity relationship feature vector between the ath rehabilitation measure and other rehabilitation measures at different measure classification levels; these four similar feature vectors are concatenated column by column to form a 4×N matrix I, where N is the number of different rehabilitation measures found from the historical rehabilitation plans, and I is the comprehensive similarity matrix of the ath rehabilitation measure and other rehabilitation measures at all levels; A rehabilitation measure interaction relationship prediction network including multiple convolutional layers and fully connected layers is used to predict the interaction relationship between rehabilitation measures; the comprehensive similarity matrix I is input into the rehabilitation measure interaction relationship prediction network, and the features of the subspace are extracted from the input tensor; three fully connected layers are set in the rehabilitation measure interaction relationship prediction network, according to the 65 categories of RRI events divided in RRI-DMI, and the interaction relationship between rehabilitation measures (A, B) can be divided into five types: A→B synergy, B→A synergy, no obvious interaction, A→B antagonism, B→A antagonism, where A and B correspond to the aforementioned a-th rehabilitation measure and b-th measure respectively; therefore, the number of neurons in the three fully connected layers is set to 256, 65 and 5 respectively; the first two fully connected layers use linear activation functions, and the last layer uses sigmoid activation function for normalized probability prediction, and the five interaction relationships between rehabilitation measures are predicted by the rehabilitation measure interaction relationship prediction network.

7. A personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment according to claim 6, characterized in that: In the rehabilitation plan acquisition module, the training of the rehabilitation measure interaction relationship prediction network uses the Adam optimization algorithm, with an initial learning rate η = 0.001, and the loss function Loss is: Among them, y q The RRI truth value of the qth interaction relationship, 0 means there is no RRI, and 1 means there is RRI: p q It represents the predicted probability of the five interaction relationships, and the result with the largest probability is selected as the RRI prediction result; The GCN edge weights for the interaction and characterization of multiple rehabilitation measures included in the rehabilitation program are initialized based on the RRI events predicted by CNN, as follows: Get a case from the data set, record the rehabilitation plan used in the case as R, and represent the rehabilitation plan R with a weighted graph; for each rehabilitation measure in the rehabilitation plan R, initialize a node in the graph to represent it; for any two rehabilitation measures in the rehabilitation plan R, if the RRI of the two rehabilitation measures predicted by CNN exists, generate a directed edge between the corresponding two nodes in the graph structure with a weight attached: Among them, v and u represent two different rehabilitation measures in the rehabilitation program R, e vu Represents the impact of rehabilitation measures v on u; uses a mask model to analyze the personalized impact of rehabilitation measures on patients; The mask model is an MLP, which receives the patient embedding vector u. The embedding vector is constructed by concatenating the two feature vectors obtained by the blood test data processing module and the chief complaint text processing module. The concatenated vector u is used to represent the patient's characteristics in a personalized way. After processing, a vector consistent with the representation dimension of a single rehabilitation measure is output, called a mask vector, and the mask vector is element-wise multiplied with the rehabilitation measure representation to obtain The details are as follows: Among them: the mask model MLP includes a hidden layer; the activation function uses the ReLu activation function, d u and d v They represent the initial representation vectors of rehabilitation measures u and v, respectively, obtained by one-hot encoding; and represents the representation vector of rehabilitation measures adjusted by the patient-individualized impact mask; Calculated and After that, they are concatenated and input into a pre-trained MLP model. and The concatenated vectors are sequentially reduced in dimension, linearly mapped, and activated to calculate an influence factor c for updating the edge weights. vu , as follows: Among them, c vu represents the impact factor, reflecting the degree of influence of the interaction between rehabilitation measures v and u on a specific patient; The impact factor c obtained through the above steps vu Includes information on the extent to which the interaction between the two single rehabilitation measures affects the specific individual patient, represents the updated edge weight, which includes the interaction strength between rehabilitation measures after the patient's personalized factors; the edge weight is updated accordingly: After completing the initialization of nodes and edges, use the GCN training method for calculation: GCN summarizes and learns node representations by recursively aggregating and transforming the feature vectors of adjacent nodes; each layer update involves message passing, message aggregation, and node representation update, and its mathematical expression is: Among them, v and u represent two different rehabilitation measures in the rehabilitation program, representing two neighboring nodes in the graph, and there is a directed edge from u to v; represents the message vector transmitted from node u to node v; is the representation of node v in the lth layer of GCN; represents the weight of the directed edge from u to v; N(v) represents the neighborhood of node v, from which node v collects information to update its aggregate message Initialized to the corresponding rehabilitation measure representation d u , d v ; The following calculation formula is used to obtain the embedded representation g of the rehabilitation program: Where: σ is the ReLu activation function.

8. A personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment according to any one of claims 1 to 7, characterized in that: In the rehabilitation program recommendation module, the embedding results u1 and u2 obtained according to formula (3) and formula (24) are concatenated to obtain the complete embedding u representing the lth patient: l , and characterization of rehabilitation measures k ,u l represents the personalized embedding vector of the lth patient, which combines the embedding information of the patient’s blood test data and the chief complaint text. k The one-hot encoding embedding of the k-th rehabilitation measure is used to represent a single rehabilitation measure; The concatenated vector [u l |d k ] is mapped to values ​​through multiple layers of fully connected layers Represents the personalized information u based on the patient l As a rehabilitation measure k Rating: H e1 =g(W e1 ·[u l |d k ]+b e1 ) (35) H e2 =g(W e2 ·H e1 +b e2 ) (36) H e3 =g(W e3 ·H e2 +b e3 ) (37) Among them, g is the activation function; W e1 , W e2 , W e3 and W eo are all weight matrices; b e1 , b e2 , b e3 , b eo is the bias vector; H e1 , H e2 and H e3 are the outputs of the hidden layers respectively.

9. A personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment according to claim 8, characterized in that: In the rehabilitation program recommendation module, an MLP model is used as the rehabilitation measure scoring model; Using BprLoss as the loss function, the rehabilitation measure scoring model is trained. The mathematical expression of the loss function L1 is: Where: σ represents the sigmoid activation function; P n represents the historical rehabilitation plan used by the nth case in the data set; j and k represent a single rehabilitation measure in the rehabilitation measure set; is the L2 regularization used; By minimizing the loss function, As large as possible, that is, for the patient in the nth record, the corresponding rehabilitation measure d j Compared with rehabilitation measures without a history of use k The higher the score, the more rehabilitation measures scoring model was obtained after the initial training, an MLP model that can perform personalized scoring of a single rehabilitation measure based on patient information; The rehabilitation program calculated by formula (34) is embedded into g, replacing the single rehabilitation measure representation d in formula (35) and sent to the rehabilitation measure scoring model completed by the initial training. The loss function is expressed as: Update iteration. If L2(W, b) is less than 0.01, stop iteration; otherwise, perform the following steps to update parameters: m t =β1m t-1 +(1-β1)g t (42) After iterative calculation and convergence conditions are met, the final rehabilitation measure scoring model is obtained, and the rehabilitation plan is scored based on the patient's personalization: H e1 =g(W1·[u i |g j ]+b1) (47) H e2 =g(W2·H e1 +b2) (48) H e3 =g(W3·H e2 +b3) (49) 10. A personalized rehabilitation program recommendation system for elderly people with motor cognitive impairment according to claim 9, characterized in that: In the rehabilitation program recommendation module, the recommendation process is as follows: The rehabilitation measure scoring model that is finally trained outputs a score for the rehabilitation plan based on the input patient embedding and historical rehabilitation plan embedding; This score reflects the patient's preference for different rehabilitation measures, thereby enabling personalized rehabilitation program recommendations.

Images